Machine Learning Robustness in Predictive Maintenance Under Adversarial Attacks

Predictive maintenance (PdM) techniques can increase industrial productivity and reduce maintenance costs by predicting the remaining useful life (RUL) of complicated machines. However, PdM systems involve industrial internet of things (IIoT) devices and machine learning (ML) algorithms, which are prone to adversarial attacks. In this work, first PdM is developed based on four ML classification models: Random Forest (RF), Light Gradient-Boosting Machine (LGBM), Gaussian Naive Bayes (GNB), and Adaptive Boosting (AdaBoost) classifier. Second, the robustness of the ML models under three adversarial attacks is evaluated, using the NASA turbofan engine dataset: Zeroth Order Optimization (ZOO), Universal Adversarial attack, and HopSkipJump attack. Results indicate RF as the most efficient classifier, reaching 96.35% of classification accuracy. Moreover, RF is proven to be the most robust of the examined classifiers under the considered attacks, displaying comparative resilience of up to 83.58% higher, compared to other models.